MODEL REFERENCE · OCTOBER 4, 2026
Choose up to three. Start with the task and access you need, then compare capability and cost.
These are dated reference profiles. Benchmark leaders refresh daily; provider links show the latest availability and pricing. An app subscription does not automatically include API credit.
A multimodal model with reasoning and tool calling. Compare accuracy and total task cost rather than token prices alone.
1M tokens. Peak uncached-input/output rates shown. Off-peak rates are $0.15 / $0.60; cached inputs cost less. See the provider schedule.
Checked 2026-10-04. Provider documentation ↗ · Benchmark source ↗
A model for long-context work, coding and vision, with low, high and max reasoning settings.
1M tokens. Uncached input/output rates; cached input is $0.30 per million tokens.
Checked 2026-10-04. Provider documentation ↗ · Benchmark source ↗
A flagship option for demanding analysis and coding. Compare its additional cost with Sonnet on work you can verify.
1M tokens. Standard API rates; fast mode and caching differ.
Checked 2026-10-04. Provider documentation ↗ · Benchmark source ↗
A lower-cost Claude option for regular coding and analysis. Its high-effort results make it worth comparing before choosing Opus.
See model documentation. Standard API rates; caching differs.
Checked 2026-10-04. Provider documentation ↗ · Benchmark source ↗
OpenAI’s highest-capability option for difficult reasoning, coding and tool-driven work.
1.05M tokens. Standard text-token rates; inputs over 272K cost more. Tools, caching and faster service have separate rates.
Checked 2026-10-04. Provider documentation ↗ · Benchmark source ↗
A less expensive alternative to Astra for coding and professional work. Test whether the difference matters on your tasks.
1.05M tokens. Standard text-token rates; inputs over 272K cost more. Tools, caching and faster service have separate rates.
Checked 2026-10-04. Provider documentation ↗ · Benchmark source ↗
A low-cost option for focused tasks, extraction and larger request volumes.
1.05M tokens. Standard text-token rates; inputs over 272K cost more. Tools, caching and faster service have separate rates.
Checked 2026-10-04. Provider documentation ↗ · Benchmark source ↗
Text and image understanding with adjustable reasoning effort. The Grok app adds search, voice and other tools beyond the model itself.
500K tokens. Base API rates up to 200K input tokens; longer context and regional processing cost more.
Checked 2026-10-04. Provider documentation ↗ · Benchmark source ↗
Google’s current Flash model for agent workflows and software engineering. Useful to compare with smaller models when latency matters.
See model documentation. Promotional standard rates through Dec 31, 2026; $1.50 input / $7.50 output from Jan 1, 2027. Grounding and caching cost extra.
Checked 2026-10-04. Provider documentation ↗ · Benchmark source ↗
A lightweight model aimed at throughput and cost-sensitive tasks. Validate extraction quality before increasing volume.
See model documentation. Standard API rates; grounding and cache storage have separate charges.
Checked 2026-10-04. Provider documentation ↗
Announced for a restricted rollout to trusted cyber defenders. Strong benchmark results are not an invitation to subscribe for access.
1M tokens. No broadly available price listed here.
Checked 2026-10-04. Provider documentation ↗ · Benchmark source ↗
An image model to consider for visual work. Compare text rendering, edits and composition using the image leaderboard and your own brief.
Image workflow. Image costs depend on dimensions and quality; text-token rates are not a useful comparison.
Checked 2026-10-04. Provider documentation ↗
Another current OpenAI image model. Use the same prompt and output size when comparing with Sunburst.
Image workflow. See current image pricing and output settings.
Checked 2026-10-04. Provider documentation ↗
The engine that generates an answer. A model’s context window and reasoning settings affect what it can process.
Search, files, connectors and editing features around that engine. Two apps using the same model can behave very differently.
A workflow that can take multiple steps and use tools. Finished-task quality depends on instructions, permissions and checks as well as the model.